daimon-memory

Persistent typed memory for AI agents. Deterministic, LLM-free recall across every tool.

---

Project status

Status quo. daimon-memory is not receiving primary development focus right now.

Its successor is dm-lite (daimon-memory v2): one binary, hybrid keyword + vector recall, and the shared memory engine for Hermes, Codex, Claude Code, and other agents. New AI memory work is focused there.

The future of this repository is undecided. It remains available and usable as documented below. If you are starting fresh, prefer dm-lite.

---

What it is

daimon-memory is a self-hostable memory backend for AI agents. It stores typed records (decisions, lessons, incidents, reminders, and more) in PostgreSQL as the canonical source, with Qdrant as a rebuildable semantic index. Recall is hybrid keyword + vector search fused by Reciprocal Rank Fusion -- no LLM in the recall path, so it is fast, cheap, and reproducible.

It is single-subject by design: one memory space per tenant, shared across every tool that connects to it. Establish a fact in Claude Code; recall it in Hermes. The same persona and operating discipline load into every agent at session start.

Full documentation: wakbijok.uk/man/daimon-memory

---

Quick start

1. Clone and configure

git clone https://github.com/wakbijok/daimon-memory && cd daimon-memory
cp .env.example .env

Open .env and set at minimum:

DAIMON_PG_PASSWORD=<something-strong>
DAIMON_API_KEY=<openssl rand -hex 32>

2. Start the stack

./install.sh

Pass --yes to accept all defaults non-interactively. The installer writes .env, brings up the Docker Compose stack, seeds the default protocols, and offers to run the persona wizard.

First start downloads the embedding model (about 130 MB) once.

3. Smoke test

# Health
curl -s localhost:8080/readyz

# Store a memory
curl -s -XPOST localhost:8080/v1/memory \
  -H 'content-type: application/json' \
  -H 'Authorization: Bearer <your-api-key>' \
  -d '{
    "kind": "decision",
    "namespace": "resources/architecture/decisions",
    "title": "Adopt Postgres + Qdrant",
    "body": "Postgres is the canonical store; Qdrant is a rebuildable vector index.",
    "fields": { "context": "needed shared memory", "rationale": "deterministic recall, rebuildable index" }
  }'

# Recall
curl -s -XPOST localhost:8080/v1/recall \
  -H 'content-type: application/json' \
  -H 'Authorization: Bearer <your-api-key>' \
  -d '{"query":"how should we store memory"}'

No API key set? Omit the Authorization header. The API is open when DAIMON_API_KEY is unset -- fine on localhost, not on a shared network.

---

Documentation

Everything beyond the smoke test lives at wakbijok.uk/man/daimon-memory:

• Installation detail (Docker Compose, Kubernetes, from source, AVX2 notes)
• Core concepts: memory kinds, namespace grammar, daimon:// URIs, the persona + discipline system layer
• Usage: the REST API, MCP tool surface, the daimon CLI (ops, backup/restore, persona authoring)
• Integration guides for Claude Code, Codex, and Hermes
• Upgrade notes and observability

---

Integrations

daimon-memory works with Claude Code, Codex, and Hermes today.

• Claude Code + Codex plugins live in their own marketplace repo: wakbijok/daimon-memory-plugins -- /plugin marketplace add wakbijok/daimon-memory-plugins.
• Hermes ships as a pip package in this repo under integrations/hermes (pip install daimon-hermes && daimon-hermes setup).
• MCP endpoint (/mcp) speaks the synchronous JSON-RPC subset of MCP, not streamable-HTTP/SSE. Hosts that require SSE can use the REST hooks at /v1 instead.
• Memories live in daimon independently of any tool, so uninstalling a tool never deletes your data.

---

Contributing

Issues and PRs welcome. cargo test --workspace should pass; please keep the core crate free of I/O and model calls (that is what makes recall deterministic). For larger changes, open an issue to discuss the design first.

License

MIT.